A couple weeks ago we wrote that OpenAI’s GPT-5.6 was here “sort of” — available only to a small group of government-approved partners while it went through federal review. On July 9 that changed: GPT-5.6 rolled out to the public, and it came as a family of three models plus a new agent built to do whole jobs. If you run a small business, the useful question is not which model wins a benchmark. It is which tier you should actually reach for, and whether the new agent is worth your attention.

Three tiers, one decision

OpenAI split GPT-5.6 into three named capability tiers. Sol is the top model, built for the hardest reasoning and long, multi-step work, priced at $5 per million input tokens and $30 per million output. Terra is the balanced everyday tier — roughly the performance of last generation’s flagship at about half the cost, at $2.50 and $15. Luna is the fast, cheap option at $1 and $6, for high-volume simple tasks. The point of the naming is that these tiers now advance on their own schedules, so “Terra” will mean “the balanced one” across future releases too.

Which one an SMB should actually use

Most small-business work lives on Terra. Drafting, summarizing, answering customer questions, cleaning up a spreadsheet — the balanced tier handles all of it well and you would be overpaying to run it on Sol. Save Sol for the genuinely hard, high-stakes jobs: a thorny contract analysis, a multi-step research task, agent work that has to hold together over an hour. Use Luna for the firehose stuff — tagging, sorting, simple classification at volume. The mistake to avoid is defaulting everything to the most powerful model; that is how you pay flagship prices for grunt work.

ChatGPT Work is the more interesting story

Alongside the models, OpenAI launched ChatGPT Work, an agent that blends its coding tool with ChatGPT to carry out full tasks across your apps and files — staying on a project for hours and turning a goal into finished output rather than just answering a question. This is the same direction Anthropic’s Cowork is heading. For an SMB, the honest framing is: this is where the real productivity is going, and also where the real oversight questions live. An agent that works for an hour unattended is powerful and needs a human checking what it produced.

The honest caveat

Do not switch your whole operation to a model that is a week old. Model names and pricing are moving fast right now — GPT-5.6 became the preferred model in Microsoft 365 Copilot the same week it launched. Test the tier you are considering on a few of your real tasks before you commit, and remember that the best model for you is often the one already sitting inside the software you use, not the one with the newest headline.

The takeaway

GPT-5.6’s public launch is worth ten minutes: try Terra on a task you do weekly, see if it beats what you use now, and keep Sol in reserve for the hard problems. Watch ChatGPT Work and Claude’s Cowork as the same bet from two directions — agents that finish jobs — but pilot them on low-stakes work first. The tiers are a gift for cost control if you match the model to the job instead of always grabbing the biggest one.